1

Automated Detection of Atrial Fibrillation using R-R intervals and multivariate based

classification.

Alan Kennedy*a, Dewar D. Finlaya, Daniel Guldenringa, Raymond R. Bondb, James

McLaughlina, Kieran Moranc.

aNanotechnology and Integrated BioEngineering Centre, University of Ulster, Northern

Ireland, UK.

bComputer Science Research Institute, University of Ulster, Northern Ireland, UK.

cDublin City University, Ireland.

Corresponding Author:

Alan Kennedy

NIBEC Building

University of Ulster, Jordanstown

Newtownabbey

Co. Antrim

BT37 0QB

2

Abstract

Automated detection of AF from the electrocardiogram (ECG) still remains a challenge. In

this study we investigated two multivariate based classification techniques, Random Forests

(𝑅𝐹) and k-nearest neighbor (𝑘 − 𝑛𝑛), for improved automated detection of AF from the

ECG. We have compiled a new database from ECG data taken from existing sources. R-R

intervals were then analyzed using four previously described R-R irregularity measurements:

(1) The coefficient of sample entropy (𝐶𝑜𝑆𝐸𝑛) (2) The coefficient of variance (𝐶𝑉) (3) Root

mean square of the successive differences (𝑅𝑀𝑆𝑆𝐷) and (4) median absolute deviation

(𝑀𝐴𝐷). Using outputs from all four R-R irregularity measurements 𝑅𝐹 and 𝑘 − 𝑛𝑛 models

were trained. RF classification improved AF detection over 𝐶𝑜𝑆𝐸𝑛 with overall specificity of

80.1% vs. 98.3% and positive predictive value of 51.8% vs. 92.1% with a reduction in

sensitivity, 97.6% vs. 92.8%. 𝑘 − 𝑛𝑛 also improved specificity and PPV over

𝐶𝑜𝑆𝐸𝑛 however the sensitivity of this approach was considerably reduced (68.0%).

1 Introduction

The automated detection and management of patients with atrial fibrillation (AF) is becoming

a priority in healthcare systems globally [1], [2]. Patients suffering from AF have a 5-fold

increase in the likelihood of suffering a stroke event [3], making early detection and

intervention a priority. Typically, when patients are suspected of having AF, a 12-lead

electrocardiogram (ECG) is recorded and this remains the gold standard for AF detection [4].

However, some AF patients suffer short and intermittent episodes which are difficult to

confirm using the 12-lead ECG, given the relatively short duration of the recording (10

seconds). Patients who are suspected of having AF, which is not confirmed on the 12-lead

ECG, are usually referred for extended (24-72 hour) ECG monitoring in an effort to detect

the suspected AF events [4]. This extended continuous monitoring is usually performed using

a Holter monitoring system although recently a number of new devices have emerged that

3

specifically focus on the detection of AF. These include disposable patch systems [5] and

iPhone based ECGs [6]. Due to the extended nature of continuous ECG recordings and the

amount of associated data generated it becomes more challenging to perform human

interpretation. Ideally, AF would be automatically detected via a software based algorithm,

alleviating the need for time consuming manual review of ECGs by clinicians. Accurate

automated detection of AF from the ECG still remains a challenge due to sustained ectopic

beats (Figure 1), motion artefact and with some devices, T-wave over sensing [7]. AF is

confirmed on the ECG as the absence of the P-wave, an irregular R-R interval and in some

patients a fibrillatory wave on the baseline ECG [8]. Most ambulatory systems rely on the

analysis of the R-R interval alone for the detection of AF. This is due to the fact that the P-

wave, and, in particular, the fibrillatory wave have a low signal-to-noise ratio during

ambulation. In this study we aim to determine if the automated detection of AF from R-R

intervals can be improved with Random Forests (𝑅𝐹) and k-nearest neighbour (𝑘 − 𝑛𝑛)

classification models.

Figure 1. An example of normal sinus rhythm (top row), atrial fibrillation (middle row) and

sustained pre-atrial contractions (bottom row).

4

2 Method

2.1 Study Data

To facilitate our experiments, we developed a new database from a range of different sources.

The database structure is outlined in the Table 1. Automated QRS detection was performed

on lead II using the open source GQRS algorithm [9], this method of QRS detection was

chosen over the manual human annotations as it allows for the automated detection of QRS

complex’s, which as described previously is highly desirable. The GQRS detection algorithm

was chosen based on its improved performance over other available QRS detection

algorithms [10]. As can be seen in Table 1 the database (𝐴𝐹𝐷𝐵!"#$) contained 322 records

consisting of a total of 4232410 automatically detected ECG beats of which 17% were

labeled as AF based on the reference rhythm annotations provided with the existing

databases. The database also contained a significant number of pre-atrial contractions (4490)

and pre-ventricular contractions (101213), the main sources of false positives for automated

AF algorithms [7]. The 322 records were then randomly split into a training dataset of 249

patients (75%) and a testing dataset of the remaining 73 patients (25%).

Table 1. The number of AF and non AF beats detected using the GQRS algorithm. Also the

number of PVC and PAC beats within each database taken from provided human beat

annotations.

5

2.2 R-R interval pre-processing

A simple three-point median filter was applied to the R-R interval data prior to analysis. This

filter is commonly implemented in an effort to remove sporadic ectopic beats from the R-R

series before AF detection is attempted [11]. The filter is defined as below:

𝑅𝑅!" = 𝑚𝑒𝑑𝑖𝑎𝑛{𝑅𝑅 𝑛 − 2 ,𝑅𝑅 𝑛 − 1 ,𝑅𝑅 𝑛 } (1)

Where 𝑅𝑅!" is the output of the median filter and 𝑅𝑅 is the time series of R-R intervals.

2.3 R-R irregularity measurements

For this study we implemented four R-R irregularity measurements (1) the coefficient of

variance (2) the root mean square of the successive differences (3) the coefficient of sample

entropy and (4) the median absolute deviation. All measurements were implemented with

moving segment window of 30 R-R intervals.

2.3.1 The coefficient of variance

The coefficient of variance (𝐶𝑉) was calculated as follows:

𝐶𝑉 = !!!!!!

(2)

Where 𝑅𝑅! is the standard deviation of the R-R interval and 𝑅𝑅! is the mean R-R interval.

Episodes of AF are expected to have a greater value of the 𝐶𝑉 than those of NSR [12].

2.3.2 The root mean square of the successive differences

The root mean square of the successive difference (𝑅𝑀𝑆𝑆𝐷) is calculated as follows:

𝑅𝑀𝑆𝑆𝐷 = !!!!

( 𝑅𝑅 !!! − 𝑅𝑅 !)!!!!!!! (3)

Where i is the R-R interval and N is the length of the segment window. Since AF exhibits

higher variability in the R-R interval than NSR, the 𝑅𝑀𝑆𝑆𝐷 is expected to be higher during

AF than NSR [13] .

6

2.3.3 The coefficient of sample entropy

The coefficient of sample entropy (𝐶𝑜𝑆𝐸𝑛) was first described by Lake and Moorman and is

based on the concept of sample entropy [14] which is calculated as follows:

𝑆𝑎𝑚𝑝𝐸𝑛 = − ln 𝑐𝑝 = − ln !!= − ln 𝐴 + ln(𝐵) (4)

Where 𝐴 is the number of matches at 𝑚 and 𝐵 is the number of matches at 𝑚 + 1, within a

tolerance of 𝑟. The coefficient of sample entropy [15] is the natural negative logarithm of the

conditional probability that R-R intervals that match at length 𝑚 will matches at length

𝑚 + 1.

𝐶𝑜𝑆𝐸𝑛 = 𝑆𝑎𝑚𝑝𝐸𝑛 + ln 2𝑟 − ln(𝑚𝑒𝑎𝑛 𝑅𝑅 ) (5)

The 𝐶𝑜𝑆𝐸𝑛 is expected to be greater during AF than normal sinus rhythm (NSR), again due

to the irregularity of the R-R interval during AF episodes.

2.3.4 The median absolute deviation

The median absolute deviation (𝑀𝐴𝐷) was popularized by Hampel who attributed its

discovery to Gauss [16]. 𝑀𝐴𝐷 is a robust measurement of the variability in a numerical

series of data. MAD is calculated as follow:

𝑀𝐴𝐷 = 𝑚𝑒𝑑𝑖𝑎𝑛( 𝑅𝑅! −𝑚𝑒𝑑𝑖𝑎𝑛 𝑅𝑅!"# ) (6)

Where 𝑅𝑅! is the R-R interval and 𝑅𝑅!"# is all the R-R intervals within the segment window.

The R-R intervals during an episode of AF have a larger variability therefore AF is expected

to have a greater value of 𝑀𝐴𝐷 than NSR [17].

2.4 Multivariate based classification methods

For this study we focused on two main classification methods, 𝑅𝐹 and 𝑘 − 𝑛𝑛. Both

supervised machine learning models were trained using the four R-R irregularity

measurements as input features and the human rhythm annotations used as an output

7

reference. These classification approaches were chosen in an effort to incorporate the best

aspects of the existing algorithms in a way that allows for better discrimination between NSR

and AF.

2.4.1 Random Forests

𝑅𝐹 is an ensemble machine learning method for classification, in this case classifying rhythm

as either NSR or AF. 𝑅𝐹 creates many classification trees from subsets of the data created

using the bootstrap method with replacement. To determine the optimal number of trees to

grow for the RF classifier the out-of-bag classification error was calculated over a number of

investigated forests sizes. The optimal number of trees grown was found to be 30.

2.4.2 K Nearest Neighbour

𝑘 − 𝑛𝑛 is a method of pattern recognition whereby new R-R intervals are classified as NSR

or AF based on the majority vote of the number of closest surrounding neighbors in feature

space. A standard 10-fold cross validation was performed on the training dataset to determine

an optimal value for 𝑘, which in this case was 17.

2.5 Statistical Analysis

2.5.1 Algorithm Training

For all four implemented R-R irregularity measurement receiver operating characteristic

(𝑅𝑂𝐶) curves were created and the area under the 𝑅𝑂𝐶 curve (𝐴𝑈𝐶) calculated. From each

𝑅𝑂𝐶 curve the optimal detection thresholds were defined as the minimum Euclidean distance

from perfect classification. Defined thresholds were then taken and applied to the testing

dataset.

2.5.2 Algorithm Testing

The performance on the four R-R irregularity measurement and two classification models

were assessed using overall beat-by-beat sensitivity, specificity and positive predictive value

(PPV). As well as this individual beat-by-beat accuracy was calculated and confidence

8

intervals created using bootstrap with replacement over 1000 bootstrap trials. Finally,

significant differences in accuracy calculations across all six detection approaches on the

testing dataset were determined using the Wilcoxon signed rank test (𝛼 = 0.05).

2.6 Results

2.6.1 Algorithm Training

When assessing the performance of the four R-R irregularity measurements on the 249

patients of the 𝐴𝐹𝐷𝐵!"#$ training dataset (results presented in Table 2), 𝐶𝑜𝑆𝐸𝑛 performed

best with AUC of 0.92 followed by RMSSD with AUC of 0.91. The ROC plots shown in

Figure 2 are provided to allow comparison between the four different automated AF detectors

on the 𝐴𝐹𝐷𝐵!"#$ and 𝑀𝐼𝑇 − 𝐵𝐼𝐻 𝐴𝐹𝐷𝐵. Outputs of each R-R irregularity measurement

were then used as input features to develop the 𝑅𝐹 and 𝑘 − 𝑛𝑛 classification models. Optimal

detection thresholds were chosen from the ROC plot of the 𝐴𝐹𝐷𝐵!"#$ (Figure 2a) and

defined as the minimum Euclidean distance from perfect classification.

9

Figure 2. ROC curves for AF detection for the four implemented R-R irregularity

measurement on (a) 𝐴𝐹𝐷𝐵!"#$ training set and the (b) 𝑀𝐼𝑇 − 𝐵𝐼𝐻 𝐴𝐹𝐷𝐵. Where (o) is the

optimal detection thresholds.

0 0.2 0.4 0.6 0.8 11-Specificity

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Sensitivity

0 0.2 0.4 0.6 0.8 11-Specificity

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Sensitivity

CoSEn

CV

RMSSD

MAD

10

Table 2. Area under the ROC curve values for the 𝑀𝐼𝑇 − 𝐵𝐼𝐻 𝐴𝐹𝐷𝐵 and the 𝐴𝐹𝐷𝐵!"#$

training dataset used in this study. The performance of the 𝐶𝑜𝑆𝐸𝑛 and 𝑀𝐴𝐷 algorithms are

reduced on the 𝐴𝐹𝐷𝐵!"#$ dataset compared to the 𝑀𝐼𝑇 − 𝐵𝐼𝐻 𝐴𝐹𝐷𝐵.

2.6.2 Algorithm and Classification Model Testing

From the 73 patients of the 𝐴𝐹𝐷𝐵!"#$ testing dataset (results presented in Table 3), RF

classification improved AF detection over 𝐶𝑜𝑆𝐸𝑛 with overall specificity of 80.1% vs. 98.3%

and positive predictive value of 51.8% vs. 92.1% with a reduction in sensitivity, 97.6% vs.

92.8%. 𝑘 − 𝑛𝑛 also improved specificity and PPV over 𝐶𝑜𝑆𝐸𝑛 however the sensitivity of this

approach was considerably reduced (68.0%). When assessing R-R irregularity approach’s

individually (Figure 3), the 𝐶𝑜𝑆𝐸𝑛 provided the best AF detection accuracy compared to the

next best algorithm 𝑅𝑀𝑆𝑆𝐷 (median accuracy, 96.9% vs. 92.7%, p < 0.001).

Figure 3. The individual accuracy of the six investigated detection algorithms. The RF model

improved the median accuracy but the difference is very small however the 5% and 95%

bootstrap confidence intervals are much smaller for 𝑅𝐹.

RF KNN CoSEn CV RMSSD MAD

Algorithm

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

%

SensitivitySpecificityPPV

11

The individual patient accuracy calculations, shown in Figure 3, demonstrated a small

improvement in median accuracy for AF detection from 𝑅𝐹 of 99.1% compared to 96.9%

achieved by 𝐶𝑜𝑆𝐸𝑛. Also of interest from these results is the clear reduction in bootstrap

confidence intervals. To determine the performance of the AF algorithms on recordings

containing ectopic beats, records from the testing dataset containing PVCs and PACs were

isolated and median AF detection accuracy calculated to establish performance.

As shown in Table 4, the 𝑅𝐹 model appeared to be more effective on both records with a

PVCs with median accuracy of 98.6% compared to 85.7% achieved by 𝐶𝑜𝑆𝐸𝑛. Also was true

with the records containing PACs with median accuracy of 99.1% compared to 94.0%

achieved by CoSEn. Although there are a limited number of recordings containing annotated

PVCs and PACs in the testing dataset of this study, results from these isolated records appear

to indicate the RF model is more effective in reducing false positives due to ectopic beats.

However further investigation is required.

Table 4. The median accuracy of 𝐶𝑜𝑆𝐸𝑛 and 𝑅𝐹 algorithm on recordings containing

annotated PVCs and PACs, the two major sources of false positives from AF algorithms,

from the testing database.

2.7 Discussion

Recently a number of studies have demonstrated improvement performance of automated AF

detection on the MIT-BIH AFDB [13], [18]–[20] . As highlighted in Figure 2, new databases

should be explored in order to develop more robust methods of automated AF detection. The

MIT-BIH database is commonly used for reporting new algorithm performance and remains

12

the most used database for algorithm testing and comparison. In this study we assembled a

large database which may be more appropriate for automated AF algorithm training and

testing. Performance of algorithm training on the training dataset is compared to the MIT-

BIH database in Figure 2. The performance of CoSEn and MAD were significantly reduced

when employing 𝐴𝐹𝐷𝐵!"#! during development in comparison to 𝑀𝐼𝑇 − 𝐵𝐼𝐻 𝐴𝐹𝐷𝐵. This

in turn highlights that results obtained using the 𝑀𝐼𝑇 − 𝐵𝐼𝐻 𝐴𝐹𝐷𝐵 may be optimistic and

not truly representative. During algorithm testing, the best performing R-R irregularity

method for diagnosing AF was the 𝐶𝑜𝑆𝐸𝑛, which provided a significant improvement in

median accuracy over the next best method the 𝑅𝑀𝑆𝑆𝐷. This observation falls in line with

the work described by Langley et al. [21] who demonstrated the improved accuracy of

𝐶𝑜𝑆𝐸𝑛 over CV and RMSSD for short R-R interval recordings. When Moorman and Lake

first described the principle of the CoSEn [14] they also demonstrated its improved

performance over CV in classifying AF from NSR. Petrenas et al. [22] investigated the

performance of 𝐶𝑜𝑆𝐸𝑛 for AF detection from the R-R interval and found that for small

segment window lengths (<30 beats) the sensitivity, specificity and accuracy of AF detection

was significantly reduced in recordings with high signal noise, also the authors demonstrated

a reduction in the performance of 𝐶𝑜𝑆𝐸𝑛 during PACs confirming observations made in [7].

The 𝑅𝐹 model described in this study improved the median accuracy over 𝐶𝑜𝑆𝐸𝑛 (99.1% vs.

96.9%, 𝑝 < 0.001). Also, as shown in Table 4, the RF model provides improved accuracy

from records containing PVCs and PACs. However, due to further testing of the developed

classification models is required to effectively establish their performance in the presence of

PVCs and PACs. This is important to note as PVCs and PACs occur frequently with the

prevalence increasing with age [23] and are a major source of false positives from automated

AF algorithms.

13

This study describes an improvement in the performance of automated AF detection

algorithms that are based on R-R interval based detection methods, through implementation

of a RF classification model. R-R interval irregularity is the most accessible ECG

characteristic for AF detection [8]. It must be appreciated though that the R-R interval based

approach does have limitations and these are rooted in the that features which are specific to

the function of the atria such as P-wave analysis, the P-R interval or fibrillatory wave are not

accounted for. However, analysis of these features is not without issue as they may be

difficult to accurately record during ambulation due to the relatively low magnitude of the P-

wave and fibrillatory wave on the ECG [24], [25].

2.8 Conclusions

Of conventional approaches the 𝐶𝑜𝑆𝐸𝑛 performed better than 𝐶𝑉, 𝑅𝑀𝑆𝑆𝐷 and 𝑀𝐴𝐷 in AF

detection from the 𝐴𝐹𝐷𝐵!"#$ database. The 𝑘 − 𝑛𝑛 classification model improved

specificity and PPV value over 𝐶𝑜𝑆𝐸𝑛 but with a substantial cost in sensitivity. The 𝑅𝐹

classification model also improved specificity and PPV with a smaller reduction in

sensitivity.

2.9 Acknowledgments

This project has received funding from the European Union’s Horizon 2020 Framework

Programme for Research and Innovation Action under Grant Agreement no.

643491. PATHway: Technology enabled behavioural change as a pathway towards better

self-management of CVD (www.pathway2health.eu). This work has also been supported by

the Department of Employment and Learning Northern Ireland, the Institution of Engineering

and Technology and the Health Informatics Society of Ireland.

References

[1] S. S. Chugh, R. Havmoeller, K. Narayanan, D. Singh, M. Rienstra, E. J. Benjamin, R.

F. Gillum, Y. H. Kim, J. H. McAnulty, Z. J. Zheng, M. H. Forouzanfar, M. Naghavi,

14

G. A. Mensah, M. Ezzati, and C. J. L. Murray, “Worldwide epidemiology of atrial

fibrillation: A global burden of disease 2010 study,” Circulation, vol. 129, no. 8, pp.

837–847, 2014.

[2] J. S. Healey, S. J. Connolly, M. R. Gold, C. W. Israel, I. C. Van Gelder, A. Capucci, C.

P. Lau, E. Fain, S. Yang, C. Bailleul, C. A. Morillo, M. Carlson, E. Themeles, E. S.

Kaufman, and S. H. Hohnloser, “Subclinical atrial fibrillation and the risk of stroke,”

N Engl J Med, vol. 366, no. 2, pp. 120–129, 2012.

[3] T. Sanna, H.-C. Diener, R. S. Passman, V. Di Lazzaro, R. A. Bernstein, C. A. Morillo,

M. M. Rymer, V. Thijs, T. Rogers, F. Beckers, K. Lindborg, and J. Brachmann,

“Cryptogenic Stroke and Underlying Atrial Fibrillation,” N. Engl. J. Med., vol. 370,

no. 26, pp. 2478–2486, 2014.

[4] K. Harris, D. Edwards, and J. Mant, “How can we best detect atrial fibrillation?,” J. R.

Coll. Physicians Edinb., vol. 42 Suppl 1, pp. 5–22, 2012.

[5] P. M. Barrett, R. Komatireddy, S. Haaser, S. Topol, J. Sheard, J. Encinas, A. J. Fought,

and E. J. Topol, “Comparison of 24-hour Holter monitoring with 14-day novel

adhesive patch electrocardiographic monitoring,” Am. J. Med., vol. 127, no. 1, pp. 95–

e11, 2014.

[6] N. Lowres, L. Neubeck, G. Salkeld, I. Krass, A. J. McLachlan, J. Redfern, A. A.

Bennett, T. Briffa, A. Bauman, C. Martinez, and others, “Feasibility and cost-

effectiveness of stroke prevention through community screening for atrial fibrillation

using iPhone ECG in pharmacies. The SEARCH-AF study,” Thromb Haemost, vol.

111, no. 6, pp. 1167–1176, 2014.

[7] S. Maddox, M. Hoskins, M. Lloyd, A. Mengistu, S. Rangaraju, L. Henriquez, B.

15

Davis, and F. Nahab, “High False Positive Rates of Atrial Fibrillation Detection

Among Stroke Patients who Receive Medtronic Implantable Loop Recorders,” Stroke,

vol. 47, no. Suppl 1, pp. A206–A206, 2016.

[8] S. Babaeizadeh, R. E. Gregg, E. D. Helfenbein, J. M. Lindauer, and S. H. Zhou,

“Improvements in atrial fibrillation detection for real-time monitoring,” J.

Electrocardiol., vol. 42, no. 6, pp. 522–526, 2009.

[9] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G.

Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. . Stanley, “PhysioBank,

PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex

Physiologic Signals,” Circulation, vol. 101, pp. 215–220, 2000.

[10] M. Llamedo and J. P. Martinez, “Qrs detectors performance comparison in public

databases,” in Computing in Cardiology Conference (CinC), 2014, 2014, pp. 357–360.

[11] A. Petr\.enas, V. Marozas, and L. Sörnmo, “Low-complexity detection of atrial

fibrillation in continuous long-term monitoring,” Comput. Biol. Med., vol. 65, pp. 184–

191, 2015.

[12] K. Tateno and L. Glass, “Automatic detection of atrial fibrillation using the coefficient

of variation and density histograms of RR and $Δ$RR intervals,” Med. Biol. Eng.

Comput., vol. 39, no. 6, pp. 664–671, 2001.

[13] S. Dash, K. H. Chon, S. Lu, and E. A. Raeder, “Automatic real time detection of atrial

fibrillation,” Ann. Biomed. Eng., vol. 37, no. 9, pp. 1701–1709, 2009.

[14] D. E. Lake and J. R. Moorman, “Accurate estimation of entropy in very short

physiological time series: the problem of atrial fibrillation detection in implanted

ventricular devices,” Am. J. Physiol. Circ. Physiol., vol. 300, no. 1, pp. H319–H325,

16

2011.

[15] M. Carrara, L. Carozzi, T. J. Moss, M. de Pasquale, S. Cerutti, M. Ferrario, D. E.

Lake, and J. R. Moorman, “Heart rate dynamics distinguish among atrial fibrillation,

normal sinus rhythm and sinus rhythm with frequent ectopy,” Physiol. Meas., vol. 36,

no. 9, p. 1873, 2015.

[16] C. Leys, C. Ley, O. Klein, P. Bernard, and L. Licata, “Detecting outliers: Do not use

standard deviation around the mean, use absolute deviation around the median,” J.

Exp. Soc. Psychol., vol. 49, no. 4, pp. 764–766, 2013.

[17] D. T. Linker, “Accurate, Automated Detection of Atrial Fibrillation in Ambulatory

Recordings,” Cardiovasc. Eng. Technol., pp. 1–8, 2016.

[18] X. Zhou, H. Ding, B. Ung, E. Pickwell-MacPherson, and Y. Zhang, “Automatic online

detection of atrial fibrillation based on symbolic dynamics and Shannon entropy.,”

Biomed. Eng. Online, vol. 13, no. 1, p. 18, 2014.

[19] J. Lee, B. A. Reyes, D. D. McManus, O. Mathias, and K. H. Chon, “Atrial fibrillation

detection using an iphone 4S,” IEEE Trans. Biomed. Eng., vol. 60, no. 1, pp. 203–206,

2013.

[20] C. Huang, S. Ye, H. Chen, D. Li, F. He, and Y. Tu, “A novel method for detection of

the transition between atrial fibrillation and sinus rhythm,” IEEE Trans Biomed Eng,

vol. 58, no. 4, pp. 1113–1119, 2011.

[21] P. Langley, M. Dewhurst, L. Y. Di Marco, P. Adams, F. Dewhurst, J. C. Mwita, R.

Walker, and A. Murray, “Accuracy of algorithms for detection of atrial fibrillation

from short duration beat interval recordings,” Med. Eng. Phys., vol. 34, no. 10, pp.

1441–1447, 2012.

17

[22] A. Petrėnas, L. Sörnmo, A. Lukoševičius, and V. Marozas, “Detection of occult

paroxysmal atrial fibrillation,” Med. Biol. Eng. Comput., vol. 53, no. 4, pp. 287–297,

2015.

[23] P. W. Macfarlane, “Comprehensive Electrocardiology,” Media, vol. 33, no. 2, p. 1792,

2011.

[24] A. Petrenas, V. Marozas, G. Jaru??evi??ius, and L. S??rnmo, “A modified Lewis ECG

lead system for ambulatory monitoring of atrial arrhythmias,” in Journal of

Electrocardiology, 2015, vol. 48, no. 2, pp. 157–163.

[25] A. Kennedy, D. D. Finlay, D. Guldenring, R. R. Bond, and J. McLaughlin, “Detecting

the Elusive P-Wave: A New ECG Lead to Improve the Recording of Atrial Activity,”

IEEE Trans. Biomed. Eng., vol. 63, no. 2, pp. 243–249, 2016.